SPNet: Structure preserving network for depth completion.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Reconstruction

Background:

• Depth completion, predicting dense depth maps from sparse inputs, is crucial for 3D scene understanding.
• Convolutional Neural Networks (CNNs) have advanced depth completion, but preserving fine structures remains challenging.

Purpose of the Study:

• To propose a novel Structure Preserving Network (SPNet) for accurate depth completion.
• To improve the preservation of intricate depth structures, including object boundaries and small features.

Main Methods:

• Developed a Multi-Scale Gradient Extractor (MSGE) using semi-fixed depthwise separable convolutions to capture structural information.
• Introduced a stable gradient Mean Absolute Error (MAE) loss (LGMAE) for enhanced structure reconstruction.
• Implemented a Multi-Level Feature Fusion Module (MFFM) to integrate spatial and semantic features for detailed depth maps.

Main Results:

• The SPNet demonstrated superior performance on the NYUv2 and KITTI datasets.
• Quantitative and qualitative evaluations confirmed the method's effectiveness in depth completion.
• The proposed techniques significantly improved the reconstruction of fine geometric details.

Conclusions:

• The SPNet effectively addresses the challenge of preserving structural details in depth completion.
• The integration of MSGE, LGMAE, and MFFM contributes to state-of-the-art depth completion performance.
• This work advances the capability of generating accurate and detailed dense depth maps from sparse inputs.